Custom meshcat contact visualizer not drawing contact forces correctly

I have a TrajectoryVisualizer object that includes a diagram and its context, used for visualizing trajectories and the contact forces along the trajectories.

class TrajectoryVisualizer:
	def __init__(self,):
		self.diagram = make_diagram_with_plant_and_scene_graph_and_visualizer_and_contact_visualizer() 
		self.meshcat # Meshcat object
		self.contact_vis  # ContactVisualizer object
		self.meshcat_vis  # MeshcatVisualizer object
        self.plant  # MultibodyPlant object

		self.context = self.diagram.CreateDefaultContext()
        self.context_meshcat = self.meshcat_vis.GetMyMutableContextFromRoot(
            self.context
        )
        self.context_contact_vis = (
            self.contact_vis.GetMyMutableContextFromRoot(self.context)
        )
        self.context_plant = self.plant.GetMyMutableContextFromRoot(self.context)

    def publish_trajectory(
        self,
        h: float,
        q_knots: np.ndarray,
        contact_results_list: List[ContactResults],
    ):
        """
        q_knots: (T + 1, n_q) array.
        contact_results_list: (T,) list.
        For 1 <= i <= T, contact_results_list[i - 1] is the contact forces at
         configuration q_knots[i].
        """
        assert len(q_knots) == len(contact_results_list) + 1


        self.meshcat_vis.DeleteRecording()
        self.meshcat_vis.StartRecording(False)

        for i, t in enumerate(np.arange(len(q_knots)) * h):
            self.context.SetTime(t)
            self.plant.SetPositions(self.context_plant)
            self.meshcat_vis.ForcedPublish(self.context_meshcat)

            if i > 0:
	            # Contact forces
                self.contact_vis.GetInputPort("contact_results").FixValue(
                    self.context_contact_vis,
                    AbstractValue.Make(contact_results_list[i - 1]),
                )
                self.contact_vis.ForcedPublish(self.context_contact_vis)

        self.meshcat_vis.StopRecording()
        self.meshcat_vis.PublishRecording()

However, I realized that if I delete the contact forces, and try to draw the the same trajectory, the contact forces do not show up anymore:

vis = TrajectoryVisualizer()
vis.publish_trajectory(h, q_knots, contact_results_list)  # drawing for the first time, ok.
vis.meshcat.Delete("contact_forces")
vis.publish_trajectory(h, q_knots, contact_results_list)  # contact forces do not show up anymore.

I have two questions:

1. Is the failure due to caching? If so, how do I fix it?
2. A possible solution is to create a new context every time I call vis.publish_trajectory(...), but my impression has always been that Context can be expensive to construct, which is why it is constructed only once in the constructor of TrajectoryVisualizer . Is my understanding correct?

I believe creating a new context would work, but want to find out the most efficient way to visualize.

Instead of calling general function meshcat.Delete("contact_forces") to clear the contact visualization, call the purpose-built contact_visualizer.Delete() instead. The ContactVisualizer owns the entire /contact_forces/... path tree in Meshcat. You are not supposed to be manually screwing with that tree outside of its control.

Specifically -- for speed, the ContactVisualizer uses an internal cache of whether or not a contact pair / surface is being shown. If you manually change the Meshcat paths that are owned by the ContactVisualizer without telling it, the visualization will no longer work correctly.

As an alternative to contact_visualizer.Delete(), another way to reset would be to ensure that the delete_on_initialization_event option is enabled in the params (it is enabled by default), and then just re-initialize the self.context each time instead of calling Delete(). This would be more future-proof against any visualization systems that gain re-initialization logic down the road, and is the best way to ensure you're starting from a clean slate each time.